Gradual Interpolation Intermediate Domain Auxiliary for Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) aims to improve the model's generalization ability by learning domain invariant representation, which has made significant progress. However, it becomes more challenging as domain discrepancies between source and target domains increase. We propose a GIIDA (Gradual Interpolation Intermediate Domain Auxiliary) method to overcome this issue. Unlike the traditional direct adaptation, our method constructs a series of intermediate domains in category consistency matching and confidence-based sample mixing up. Thereby learning rich interdomain information through self-training. We introduce a clean noise classifier to estimate the noise transition matrix. The clean classifier is used to assign pseudo-labels for interdomain samples. Then, the intermediate domain samples provide the optimal parameters for the noise classifier as a closed solution. In the end, we designed an uncertainty-guided regularization to alleviate the issue of overconfidence. Extensive experimental results on three common benchmarks show the effectiveness of our method while bringing remarkable improvements against the baseline.